search

IceClear / StableSR

[IJCV2024] Exploiting Diffusion Prior for Real-World Image Super-Resolution
https://iceclear.github.io/projects/stablesr/
Other
2.17k stars 139 forks source link

Details about test set construction #65

Open k-zha14 opened 1 year ago

k-zha14 commented 1 year ago

Hi, guys! Thanks for your meaningful and interesting works! I am a newbie in this field and encounter some puzzles in test set construction.

As stated in Sec. 4.2, 'Specifially, for images from[5, 29, 68], we crop them at the center to obtain patches with 128x128 resolution.'([5]->RealSR, [29]->DPED-iphone, [68]->DRealSR), I think the 128x128 crop operation should be applied at the LR images, which is paired with 512x512 operation on HR images. Is it right?

BTW, I notice that both RealSR and DRealSR have provided the X2, X3 and X4 datasets. I wonder whether the aforesaid operation only acted on the X4 datasets.

As for DPED-iPhone dataset, these are total 29 x 3 = 87 High-Resolution test images from 3 devices, i.e. blackberry, iphone and sony. My questions are: 1) are the total 87 images used? 2) we treat these images as the LR, and just crop the 128x128 center patch from these high-resolution images?

Thanks for you help!

IceClear commented 1 year ago
  1. Correct.
  2. We use x4 data for 128 to 512.
  3. (1) We use all the test images. (2) Correct.